1 . 1 I N T RO D U C T I O N
This chapter presents series on top incomes and top wages in India between the years 1922 and 2000 based on individual tax returns data. We use tabulations of tax returns published each year by the Indian tax administration to compute the share of the top percentile of the distribution of total income, the top 0.5 per cent, the top 0.1 per cent, and the top 0.01 per cent. We do the same for the wage distribution. We do not go below the top percentile because incomes below this level are largely exempt from taxation in India.
Our series begin in 1922, when the income tax was created in India, and allow us to look at the impact of the Great Depression and the Second World War on inequality. We are particularly interested in the period starting in the 1950s, right at the beginning of India’s experiment with socialism. This experiment was officially suspended in 1991 with the beginning of the liberalization process, which continued through the 1990s. One explicit goal of the socialist programme was to limit the economic power of the elite, in the context of a mixed economy.
Our data offer us the opportunity to say something about the extent to which this programme, with all its well-known deficiencies, succeeded in its distributional objectives. This is important first, because it is a vital part of our assessment of this period. And second, because it offers a window into the broader question of the role of policy in affecting the distribution of income and wealth in a developing country. Given that much of the economic activity in these countries is outside the formal sector, it is not at all obvious that there is a lot that policy can affect.1
Our results are consistent with an important role for policy in shaping the distribution of income. In particular, we do find evidence of a substantial decline in the share of the elite during the years of socialist planning and a comparable
We are grateful to Tony Atkinson, Amaresh Bagchi, Gaurav Datt, Govinda Rao, Martin Ravallion, T. N. Srinivasan, Suresh Tendulkar, and two anonymous referees for useful discussions, to Sarah Voitchovsky for excellent research assistance, and to the MacArthur Foundation for financial support.
A shorter version of this chapter was published as A. Banerjee and T. Piketty, ‘Top Indian Incomes, 1922 2000’,World Bank Economic Review, 19 (2005): 1 20.
1 Especially tax policy.
recovery in the post-liberalization era. However the rebound seems to start significantly before the official move towards liberalization.
Given that these results are likely to be controversial, it is worth emphasizing that there are a number of obvious problems with using tax data, not the least because of tax evasion. We discuss these at some length in section 1.4. While we conclude that our results are probably robust, we do not intend them to be definitive. Our view is rather that they provide a point of departure on an important question about which very little is known, primarily because of data limitations. There are good reasons to suspect that the usual sources of informa-tion on income distribuinforma-tion in India—such as consumer expenditure surveys—
are not particularly effective at picking up the very rich. This is in part because the rich are rare, and in part because they are much more likely to refuse to cooperate with the time-consuming and irksome process of being subjected to a consumer expenditure survey.2
While there is no hard evidence that the rich are indeed being undercounted in India (the Indian consumer expenditure surveys do not, for example, report refusal rates by potential income category), one reason to suspect that this is the case comes from what has been called the Indian growth paradox of the 1990s. According to the standard household expenditure survey conducted by the National Sample Survey (NSS), real per capita growth in India during the 1990s was fairly limited. Such a conclusion stands in sharp contrast with the substantial growth measured by national accounts statistics (NAS) over this same period.
This puzzle has attracted quite a lot of attention during recent years3 and it has been widely suggested that it might simply be that a very large part of the growth went to the very rich. However there has been no attempt to directly quantify this possibility.4Our data allow us to take a useful step in this direction.
We are able to put bounds on the extent to which the growth gap can be explained simply in terms of undercounting the very rich. We conclude that it can explain between 20 per cent and 40 per cent of the puzzle. Although this is not negligible,
2 See, e.g., Szekely and Hilgert (1999), who look at a large number of Latin American household surveys and find that the ten largest incomes reported in surveys are often not very much larger than the salary of an average manager in the given country at the time of survey. For a systematic comparison of survey and national accounts aggregates in developing countries, see Ravallion (2001).
3 See, e.g., Datt (1999), Ravallion (2000), World Bank (2000), Sundaram and Tendulkar (2001).
Recently released data from the 1999 2000 NSS round have revealed that NSS growth was larger than expected during the 1990s and that poverty rates did decline over this period, contrarily to what most observers believed on the basis of pre 1999 2000 NSS rounds (see Deaton and Dre`ze 2002 and Deaton 2003a, 2003b). However the overall NSS NAS growth gap still appears to be substantial, even after this correction (see Table 1.2 below), and this substantial gap remains to be explained. The existence of a discrepancy between NSS and NAS statistics was already a subject of enquiry in India during the 1980s (see, e.g., Minhas 1988 and Minhas and Kansal 1990), but the gap observed during the 1990s appears to be substantially larger than during previous decades. For a broader, international perspective on the survey vs. national accounts debate, see Deaton (2003c).
4 Sundaram and Tendulkar (2001) find that the NSS NAS gap is particularly important for commodities that are more heavily consumed by higher income groups, thereby providing indirect evidence for the explanation based on rising inequality.
2 Top Indian Incomes, 1922–2000
this leaves the bulk of the puzzle unaccounted for, largely because the share of the rich in total income is still relatively small. This suggests that there probably is some deeper problem with the way either the NSS or the NSO (which generates the NAS) collects its data.5
The rest of this chapter is organized as follows. Section 1.2 briefly outlines our data and methodology. Section 1.3 presents our long-run results. Section 1.4 discusses potential problems with this evidence. Section 1.5 uses this evidence to shed some light on the Indian growth paradox of the 1990s. Section 1.6 con-cludes.
1 . 2 DATA A N D M E T H O D O LO G Y
The tabulations of tax returns published each year by the Indian tax administra-tion in the ‘All-India Income-Tax Statistics’ (AIITS) series constitute the primary data source used in this chapter. The first year for which we have income data is 1922–3 while the last is 1999–2000.6
Due to the relatively high exemption levels, the number of taxpayers in India has always been rather small. The proportion of taxable tax units was around 0.5 per cent–1 per cent from the 1920s to the 1980s, and it rose sharply during the 1990s up to 3.5 per cent–4 per cent at the end of the decade, following the large increase in top nominal incomes (see Figure 1.1).7Therefore our long-run series cannot go below the top percentile.
5 See Bhalla (2002) for a negative view of the NSS approach. For more balanced discussions of the relative merits of survey and national accounts aggregates in developing countries, see Ravallion (2001) and Deaton (2003c).
6 All references to the relevant AIITS publications are given in Table 1A.1. Financial years run from 1 April to 31 March in India (1922 3 refers to the period running from 1 April 1922 to 31 March 1923, etc., and 1999 2000 to the period running from 1 April 1999 to 31 March 2000). Note also that AIITS publications always refer to assessment years (AY), i.e. years during which incomes are assessed, while we always refer to income years (IY) (IY AY 1). For instance, AIITS 1923 4 contains the data on IY 1922 3, etc., and AIITS 1999 2000 contains the data on IY 1998 9. AIITS 2000 1 (IY 1999 2000) was not yet available when we revised this paper, and our IY 1999 2000 figures for top incomes were obtained by inflating the 1998 9 figures by the nominal 1999 2000/1998 9 per tax unit national income growth rate. This approximation probably leads us to underestimate top income growth. We did this because there was no large NSS round for 1998 9 so it was easier to make comparison with 1999 2000 as the end point.
7 Throughout the chapter, ‘tax units’ should be thought of as individuals (all of our estimates have been obtained by summing up tax returns filed by individuals and those filed by ‘Hindu undivided families’ (HUF); the latter make less than 5% of the total in the 1990s, down from about 20% in the inter war period). The total, theoretical number of tax units was set to be equal to 40% of the total population of India throughout the period (see Table 1A.1, col. (2)). This represents a rough estimate of the potential ‘positive income population’ of India: this is lower than India’s adult population (the 15 year and over population makes up about 60 5% of total population since the 1950s), but is very close to India’s labour force (the labour force consists of about 40 5% of total population since the 1950s).
Abhijit Banerjee and Thomas Piketty 3
The tabulations published in AIITS report the number of taxpayers and the total income reported by these taxpayers for a large number of income brackets.
By using standard Pareto extrapolation techniques we computed for each year the average incomes of the top percentile (P99–100), the top 0.5 per cent (P99.5–
100), the top 0.1 per cent (P99.9–100), and the top 0.01 per cent (P99.99–100) of the tax unit distribution of total income, as well as the income thresholds P99, P99.5, P99.9, and P99.99 and the average incomes of the intermediate fractiles P99–99.5, P99.5–99.9, and P99.9–99.99.8
To get a sense of the orders of magnitude, we report in Table 1.1 the results obtained for 1999–2000. There were almost 400 million tax units in India (396.4 million). Based on the national accounts statistics, the average income of those 400 million tax units was around Rs 25,000 per year ($3,000 in PPP terms).9To
8 The Pareto law is given by 1 F(y) (k/y)a(where 1 F(y) is the fraction of the population with income above y, and k>0 and a>1 are the structural Pareto parameters). For a recent use of Pareto extrapolation techniques with similar tax return data, see Piketty (2003) and Piketty and Saez (2003).
See also Atkinson (2007; chapter 4 in Volume I) and Dell (2007; chapter 9 in Volume I).
9 Our average income series (see Table 1A.2, col. (7)) was set to be equal to 70% of national income per tax unit (the 30% deduction is assumed to represent the fraction of national income that goes to undistributed profits, non taxable income, etc.; the national income series was taken from Sivasu bramonian 2000, from whom we also took our population series). We also report in Table 1A.1 other income aggregates based on GDP and NAS household consumption (both taken from the World Bank’s WDI database, from which we also extracted our CPI series, as well as the PPP exchange rate used in Table 1.1) and on NSS household consumption (computed from Datt 1997, 1999, for the 1956 98 series and Deaton and Dre`ze (2002: n. 24) for the corrected 1999 2000/1993 4 growth rate).
0.0%
0.5%
1.0%
1.5%
2.0%
2.5%
3.0%
3.5%
4.0%
1922–3 1927–8 1932–3 1937–8 1942–3 1947–8 1952–3 1957–8 1962–3 1967–8 1972–3 1977–8 1982–3 1987–8 1992–3 1997–8
Figure 1.1 The proportion of taxable tax units in India, 1922 2000
Source: Authors’ computations using tax returns data (see Table 1A.1, col. (4)).
4 Top Indian Incomes, 1922–2000
Table1.1TopIndianincomesin19992000 ThresholdsIncomelevel (Rs) Incomelevel (US$) (market exhangerate) Incomelevel (US$)(PPP conversion factor)FractilesNumberoftax unitsAverageincome (Rs) Averageincome (US$)(market exchangerate)
Averageincome (US$)(PPP conversionfactor) (1)(2)(3)(4)(5)(6)(7)(8)(9) FullPopulation396,400,00025,6705962,968 P9987,6332,03510,131P9999.51,982,00098,8422,29511,427 P99.5147,5463,42717,057P99.599.91,585,600216,9295,03825,079 P99.9295,1036,85334,116P99.999.99356,760590,48813,71368,264 P99.991,383,93032,140159,992P99.9910039,6404,034,28993,690466,392 Source:Table1A.2andTable1A.3,row1999–00.Amountsin$havebeencomputedbyapplyingtheaverage1999–2000marketexchangerate(thatis,1$¼43.06Rs)andtheaverage 1999–2000PPPconversionfactor(thatis,1$¼8.65Rs)toamountsincurrent1999–2000Rs.
belong to the top percentile (P99), which includes about 4 million tax units, one needed to make more than Rs 88,000 (around $10,000 at PPP). The average income of the bottom half of the top percentile (fractile P99–99.5, about 2 million tax units) was about Rs 99,000 (less than $12,000 at PPP). To belong to the top 0.01 per cent (about 40,000 tax units), one needs to make more than Rs 1.4 million ($160,000 at PPP), and the average income above that threshold was more than Rs 4 million ($470,000 at PPP).10
As in other countries, the top of India’s income distribution appears to be very precisely approximated by the Pareto structural form.11 On the other hand the estimates for the recent period are subject to sampling error: the AIITS tabula-tions were based on the entire population until the early 1990s (as in most OECD countries),12 but they now seem to be based upon uniform samples of all tax returns. Although there is uncertainty about the new sampling procedure, the sampling rate seems to be sufficiently large to guarantee that the estimated trends for top income shares are statistically significant.13
AIITS publications also include tabulations reporting the amounts of the various income categories (wages, business income, dividends, interest, etc.) for each income bracket. In particular, AIITS offers separate tables for wage earners who are by far the largest subgroup. This allowed us to separate estimates for top wage fractiles, which we can compare to our top fractiles estimates for total income (see below).14
10 In order to put these numbers in global perspective, one can note that India’s 1999 2000 P99.99 threshold (about $160,000 in PPP terms) is located midway in between US 1998 P95 and P99 thresholds for 1998 (resp. $107,000 and $230,000; see Piketty and Saez (2003: table 1)), and that India’s 1999 2000 P99.9 threshold (about $34,000 in PPP terms) is well below US 1998 P90 threshold ($82,000).
11 In the same way as for other countries (see above for references), we checked that our extrapo lation results are virtually unaffected by the choice of extrapolation thresholds used to estimate the structural parameters. Pareto coefficients are locally very stable in India, just as in other countries.
Prior to the 1990s, the fraction of individuals subject to tax was less than 1%, and we used the lowest threshold available in order to estimate the top percentile threshold P99 (given that Pareto coefficients are in practice very stable, the resulting estimates appear to be as precise as estimates for thresholds P99.5 and above).
12 Or on stratified samples with sampling rates close to 100% for top incomes.
13 According to the tax administration statistics division, the sampling rate is about 1% and approximately uniform (no precise information about sampling design and rate is included in AIITS publications). Given India’s large population, this implies that our estimate for the top 1%
income share (8.95% of total income in 1999 2000) has a standard error of about 0.04%, and that our estimate for the top 0.01% income share (1.57% of total income in 1999 2000) has a standard error of about 0.08%. There is some evidence however that the sampling design is changing and that published tabulations are becoming more volatile by the end of the period. In particular, the tabulations for IY 1997 8 (AIITS 1998 9) contain far too many individual taxpayers above 1 million Rs, thereby suggesting that something went wrong in the sampling design during that year. The 1997 8 estimates were corrected downwards on the basis of 1996 7 and 1998 9 tabulations.
14 Published wage tabulations for IY 1996 7 and 1997 8 appear to suffer from sampling design failures (top wages are clearly truncated in 1996 7, and they are too numerous in 1997 8), and our estimates for those two years were corrected on the basis of 1995 6 and 1998 9 data.
6 Top Indian Incomes, 1922–2000
1 . 3 T H E LO N G - RU N DY NA M I C S O F TO P I N C O M E S H A R E S , 1 9 2 2 – 2 0 0 0
Figure 1.2 illustrates the basic pattern of our findings. Our results show that income inequality (as measured by the share of top incomes) has followed a U-shaped pattern over the 1922–2000 period. The top 0.01 per cent income share was fluctuating around 2–2.5 per cent of total income from the 1920s to the 1950s. It then gradually fell from about 1.5–2 per cent of total income in the 1950s to less than 0.5 per cent in the early 1980s, and finally rose during the 1980s–1990s, back to 1.5–2 per cent during the late 1990s. What this means is that the average top 0.01 per cent income was about 150–200 times larger than the average income of the entire population during the 1950s. It went down to less than 50 times as large in the early 1980s, but went back to being 150–200 times larger during the late 1990s.
The exact turning point is also of some interest. We see that the decline in the share of the top 0.01 per cent is relatively rapid till 1974–5. Then it slows consider-ably but there is still a clear downward trend till 1980–1. Then it reverses: the trend is upwards throughout the 1980s, reaching a peak in 1988–9. Over the 1980s, the share of the top 0.01 per cent more than doubles—from less than 0.4 per cent to more than 0.8 per cent. But it then reverses once again, and by 1991–2 it is back below 0.6 per cent. Then it takes off and after 1995–6 remains in the 1.5–2 per cent range.
One also observes a similar (though less pronounced) U-shaped pattern for the top 1 per cent income share, which went from about 12–13 per cent during the 1950s to 4–5 per cent in the early 1980s to 9–10 per cent in the late 1990s (see Figure 1.4).
0.0%
0.5%
1.0%
1.5%
2.0%
2.5%
3.0%
3.5%
1922–3 1927–8 1932–3 1937–8 1942–3 1947–8 1952–3 1957–8 1962–3 1967–8 1972–3 1977–8 1982–3 1987–8 1992–3 1997–8
Figure 1.2 The top 0.01% income share in India, 1922 2000
Source: Table 1A.5, col. (4).
Abhijit Banerjee and Thomas Piketty 7
0.0%
1.0%
2.0%
3.0%
4.0%
5.0%
6.0%
7.0%
8.0%
1922–3 1927–8 1932–3 1937–8 1942–3 1947–8 1952–3 1957–8 1962–3 1967–8 1972–3 1977–8 1982–3 1987–8 1992–3 1997–8
Figure 1.3 The top 0.1% income share in India, 1922 2000
Source: Table 1A.5, col. (3).
4.0%
5.0%
6.0%
7.0%
8.0%
9.0%
10.0%
11.0%
12.0%
13.0%
14.0%
15.0%
16.0%
17.0%
18.0%
19.0%
1922–3 1927–8 1932–3 1937–8 1942–3 1947–8 1952–3 1957–8 1962–3 1967–8 1972–3 1977–8 1982–3 1987–8 1992–3 1997–8
Figure 1.4 The top 1% income share in India, 1922 2000
Source: Table 1A.5, col. (1).
8 Top Indian Incomes, 1922–2000
Once again the turning point seems to be around 1980–1, and over the 1980s, the share of the top 1 per cent also doubles. Then, as with the share of the top 0.01 per cent, there is a period of retrenchment that lasts till 1991–2, followed by a renewed upward movement.
The comparison of Figures 1.2 and 1.3 reveals another intriguing fact: While in the 1980s the share of the top 1 per cent increases almost as quickly as the share of the top 0.01 per cent, in the 1990s there is a clear divergence between what is happening to the top 0.01 per cent and the rest of the top percentile. To confirm that this is the case, we break up the top percentile into four groups: those between the 99th percentile and the 99.5th percentile, those between the 99.5th percentile and the 99.9th percentile, those between the 99.9th percentile and the 99.99th percentile, and those in the top 0.01 percentile. Table 1.2 reports what happened to each of these groups in the 1987–2000 period. We see that only those in the top 0.1 per cent enjoyed income growth rates faster than the growth rate of GDP per capita. This contrasts with what we see when we look at the period that includes the 1980s (see Table 1.3). For this period we see evidence of above-average growth for the entire top percentile.
While 1980–1 was clearly the year when the data series turn around, it is not possible to date the ‘true’ turnaround with quite so much precision, because the share of the rich is also affected by short-run, cyclical factors. It is possible that our data put the turning point in 1980–1 only because we have not made any allowances for the deep recession of 1979–80 and 1980–1, which hurt the rich. As a result, we see a sharp upward trend starting in 1981, even though perhaps what is really happening Table 1.2 Top income growth in India during the 1990s: 1999 2000 vs. 1987 1988
1999 2000 vs. 1987 8 1999 2000 vs. 1987 8 (nominal growth) (real growth)
Household consumption/capita (NSS) þ242% þ19%
GDP/capita (NAS) þ337% þ52%
Household consumption/capita (NAS) þ304% þ40%
National income/tax unit (NAS) þ346% þ55%
Top income fractile P99 100 (tax returns) þ392% þ71%
Top income fractile P99.5 100 (tax returns) þ412% þ78%
Top income fractile P99.9 100 (tax returns) þ548% þ125%
Top income fractile P99.99 100 (tax returns) þ1009% þ285%
Top income fractile P99 99.5 (tax returns) þ331% þ50%
Top income fractile P99.5 99.9 (tax returns) þ317% þ45%
Top income fractile P99.9 99.99 (tax returns) þ393% þ71%
Top income fractile P99.99 100 (tax returns) þ1009% þ285%
Consumer price index þ188%
Share of growth gap accounted for by P99 100 20.1%
Share of growth gap accounted for by P99.5 100 17.2%
Share of growth gap accounted for by P99.9 100 12.7%
Share of growth gap accounted for by P99.99 100 8.0%
Source: Authors’ computations using tax return, NAS and NSS data (see Table 1A.2, Table 1A.3, and Table 1A.4, row 1999–2000/1987–8).
Abhijit Banerjee and Thomas Piketty 9
in 1981–2 and 1982–3 is just a reversion to the pre-existing trend. Therefore rather than naming a single year, we date the turnaround to the early to mid 1980s.
The fact that the turning point is so early makes it hard to attribute it to the formal
The fact that the turning point is so early makes it hard to attribute it to the formal